The Partially-Observable Markov-Decision-Process (POMDP) model explicitly models the uncertainty

1
The Hidden Information State (HIS) Dialogue System
Steve Young, Jost Schatzmann, Blaise Thomson, Hui
Ye, and Karl Weilhammer Cambridge University
Engineering Department sjy, js532, brmt2,
hy216, kw278_at_eng.cam.ac.uk
Statistical Approaches
The POMDP Approach
The HIS Model
The Demo System

• The Partially-Observable Markov-Decision-Process
  (POMDP) model explicitly models the uncertainty
• POMDP policy is a mapping from
• n-dimensional belief space to actions

• A statistical approach to dialogue system design
  offers many advantages
• models variance in user behaviour
• formalises design criteria as objective reward
  functions
• learn optimal policy from dialogue data
• reduces authoring costs
• potential to adapt online
• The Markov-Decision Process model views dialogue
  policy as a mapping from dialogue states to
  system actions

• The HIS model is the first trainable and scalable
  implementation of a spoken dialog system based on
  the POMDP model
• partition state space
• compute partition beliefs not state beliefs
• use grid-based approximations, hence finite
  policy table
• represent user goals by branching-tree
• incorporate user and observation model in belief
  updates

• A working prototype system has been built for the
  Tourist Information Domain
• The system was extensively tested by human users
  as part of the EU Talk Project
• 40 subjects, four dialogues each
• mix of native and non-native speakers
• Dialogue performance was measured based on
  recommendation of a correct venue
• successful task completion gt 90
• avg. turns to task completion 5.59 turns

The MDP Model
Acknowledgements
This research was partly funded by the EU FP6
TALK Project, http//www.talk-project.org